Elon Musk’s goal for xAI’s Grok model is to be “maximally truth-seeking.” When Grok generated responses that were not aligned with Musk’s ideas of Truth, he promised to “fix” Grok, which appears to involve tweaking its system prompt. The results included Grok calling itself MechaHitler after being made less ‘politically correct’. Problematic.
But is it even possible to build a truth-seeking AI?
LLMs are probabilistic machines. They predict the next token based on patterns from a massive corpus of Internet text.
When xAI added “don’t shy away from politically incorrect claims” to Grok’s prompt, they weren’t accessing Truth but adjusting probability distributions and nudging the bot’s behavior into problematic spaces.
Training Grok4 reportedly cost almost half a billion dollars. It was so expensive because model capabilities grow with the size (parameters) and the amount and diversity of the data used to train the model. LLM capabilities are an emergent behavior driven by the amount of data used to train the model.
LLMs are “grown, not crafted”. Trying to ensure that an LLM becomes “maximally truth-seeking” is a Sisyphean task.
You could train an LLM only on data that is politically acceptable – oh sorry – certified to be True. Musk is, of course, building Grokipedia – guaranteed to be free of bias and presumably used as a corpus for training “the son of Grok”.
Good luck with the benchmarks!
Elon Musk has a phenomenal track record, but he will fail to build a maximally truth-seeking AI. LLMs operate in a probabilistic world. They are phenomenally capable black boxes for which we have no coherent theoretical framework to explain their behaviors.
Tweaking the system prompt, or tweeting angrily, may nudge LLM behavior, but with unpredictable and potentially undesirable outcomes. Instead of engaging in an endless culture war, it might be more prudent to use engineering resources to develop better guardrails on LLM behavior and take a realistic assessment of current AI capabilities.